Infinite-Dimensional Feature Interaction

Authors: Chenhui Xu, FUXUN YU, Maoliang Li, Zihao Zheng, Zirui Xu, Jinjun Xiong, Xiang Chen

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments reveal that Infi Net achieves new state-of-the-art, owing to its capability to leverage infinite-dimensional interactions, significantly enhancing model performance.
Researcher Affiliation Collaboration Chenhui Xu1,2 Fuxun Yu1,3 Maoliang Li4 Zihao Zheng4 Zirui Xu1 Jinjun Xiong2, Xiang Chen1,4, 1George Mason University 2University at Buffalo 3Microsoft 4Peking University
Pseudocode No The paper provides mathematical formulations but no clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper states that datasets are open-source but does not provide any link or explicit statement about the availability of the authors' own source code for their methodology.
Open Datasets Yes We conduct image classification experiments on Image Net-1K [9], which contains 1.28 million training samples belonging to 1000 classes and 50K samples for validation. We train the Infi Net-T/S/B/L models for 300 epochs with Adam W [28] optimizer. We evaluate our models for object detection tasks on widely used MS COCO [23] benchmark. We evaluate our models for the semantic segmentation task on widely used ADE20K [48] benchmark covering 150 semantic categories on 25K images, in which 20K are used for training.
Dataset Splits Yes Image Net-1K [9], which contains 1.28 million training samples belonging to 1000 classes and 50K samples for validation.
Hardware Specification Yes Image Net-1K experiments are conducted on 4 Nvidia A100 GPUs and Image Net-21K on 16 .
Software Dependencies No The paper mentions optimizers (Adam W) and schedulers (cosine learning rate) but does not provide specific version numbers for software libraries or frameworks (e.g., PyTorch version, CUDA version).
Experiment Setup Yes We train the Infi Net-T/S/B/L models for 300 epochs with Adam W [28] optimizer. We use the cosine learning rate scheduler [27] with 20 warmup epochs and the basic learning rate is set as 4 10 3. The training resolution is set as 224 224. More details can be found in Appendix A.1. (Tables 4 and 5 in Appendix A.1 provide extensive hyperparameter details including batch size, weight decay, label smoothing, and data augmentation settings.)